The Chokepoint Rotation

The Thesis

A gigawatt is a bill of materials, and the people on it are fighting over the split.

This site tracks four binding inputs — the money loop, electricity, memory and the fabs underneath both — plus the demand all four are underwritten by. Each of those pages asks whether a constraint holds. This one asks a different question, and it is the one that decides which shares you can actually own: when the buildout spends a dollar, who keeps it?

The answer is not fixed, and the record shows it moving. Every technology gets pushed until some part of it becomes the barrier, and whoever owns that part collects the surplus until it is relieved — at which point the barrier, and the money, moves somewhere else. Lithography was the barrier for a decade. Then leading-edge foundry capacity. Then power. Right now it is memory, and the numbers on the memory side of the chain are not in the same league as anyone else’s.

What makes this a civil war rather than a rotation is that the pie is fixed at the point of sale. A hyperscaler signs for a gigawatt at a price. Everything downstream of that number — wafers, EUV passes, HBM stacks, substrates, optics, transformers — is a claim on the same dollar. When the memory makers took more of it in 2026, they did not take it from the customer. They took it from the other links.

The record’s clearest evidence is not an analyst’s estimate. It is four price increases, announced along one chain, inside four months, each one aimed at the customer immediately above.

The Evidence

Rank the chain by how fast profit grew in each layer’s most recent reported quarter, and the ladder is unmistakable.

Memory sits at the top and it is not close. Samsung reported operating profit up 1,814% on memory demand. SK Hynix reported operating profit up 557%. Micron reported fiscal Q3 revenue up 346%, a single quarter larger than its entire prior fiscal year. Then a gap. Nvidia, the company the whole trade is named after, reported revenue up 106% and net income up 126%. Then TSMC, which makes the chips Nvidia designs: revenue up 36%, net income up 77.4%. Then, at the bottom, the company with the most complete monopoly in the entire chain — ASML, whose Q2 net sales of €9.3B against €7.7B a year earlier is growth of about 21%.

That inversion is the finding. ASML has the strongest structural position in semiconductors — one supplier, no alternative, a machine nobody else can build — and it is the slowest-growing link in the chain it makes possible. Durability and this cycle’s increment are different things. A monopoly on a step that is already installed collects a toll; a shortage on a step nobody built enough of collects a ransom.

The market has noticed. In June, Bloomberg reported ASML’s stock up 64% year to date but trailing the US chip sector as the industry spent more on processes outside lithography — etching, deposition, advanced packaging. The same mechanism appears in the cost structure: on the compute-economics read this site keeps, lithography is only in the teens as a share of a DRAM wafer’s cost against roughly 30% for logic. The AI cycle is disproportionately a memory-and-packaging cycle, and lithography’s share of it is smaller than its reputation.

Then the chain started repricing itself, link by link. In July, sources said ASML had discussed raising EUV prices with TSMC and planned to charge 10% more for DUV systems — and that TSMC was resisting. Also in July, Samsung raised prices up to 15% on advanced 4nm and 5nm contract manufacturing for new orders, explicitly on AI demand and TSMC’s tight capacity. The memory makers had already moved, using their leverage to lock buyers into long-term agreements in a market where the Wall Street Journal said AI had made memory more valuable than oil. And in August, Nvidia’s top customers were told prices will jump 15% or more on Vera Rubin and Grace Blackwell systems from early 2027.

Read those four together and the shape is plain. Equipment raises on foundry. Foundry raises on designers. Memory raises on everyone. Designers raise on the labs and the clouds. Nobody in the middle is absorbing anything; each link is passing the increase up and keeping what it can. The last buyer in the line has no one to pass it to, which is why every one of these increases eventually lands on the same token demand curve — and why the year that curve arrives decides whether the whole arrangement was clever or ruinous.

The tell that memory has taken too much is Nvidia designing it out. In August, Nvidia was reported to be weighing lower-memory versions of its Rubin Ultra GPU because it may not secure enough HBM, having tested at least three configurations. The most powerful buyer in the chain, considering shipping a worse product rather than pay what the input costs, is the clearest evidence available that the split has moved past what the buyer thinks fair. It is also the mechanism by which the rotation continues: the barrier gets engineered around, and the surplus moves on.

And the buyer is spending to make sure the layer above it never gets its own turn. Nvidia agreed to buy the open-model repository Hugging Face for $12.9B, and is reportedly using a $6B deal with Poolside to build an open-weight model to compete with DeepSeek and Kimi. It put its Groq 3 LPX inference accelerator into full production with Nebius as first customer, after licensing Groq’s technology and hiring its team — a deal that left Groq’s next round at $3.5B, down from $6.9B. And it led the sector in startup investing, part of $250B+ of chip-adjacent rounds year to date. That is what the open-weight question looks like from the silicon layer: not a threat, but a purchase. Every model given away is a model that cannot charge rent on the layer above the chip, which leaves more of the split for the chip. See the open weight trade for the same moves read from the model side.

The Companies

ASML holds the most durable position and this cycle’s weakest hand — a combination its dossier has to hold in one frame rather than resolving. Its Q4 bookings of €13.2B beat estimates by more than double, and Intel committed to High NA EUV for Panther Lake, so the order book is not the problem. What the record shows is a monopolist whose customers are resisting a 10% increase while spending freely on the steps around it, in a cycle whose bottleneck is somewhere else. The export-control exposure the dossier tracks — including US questions about an EUV machine reaching China, which ASML called inaccurate — is a separate risk that runs through the China AI stack.

TSMC is the junction, and its position is visible in what it refuses. It is resisting ASML’s price plan from below while Samsung raises its own prices citing TSMC’s tight capacity from the side. Its quarter — revenue up 36%, net income up 77.4%, with chips 7nm or smaller at 77% of wafer revenue, up from 74% a quarter earlier — is a company converting mix rather than volume into margin. That is the profile of a link with real power that is nonetheless not where the shortage is.

Marvell is the custom-silicon and interconnect claim on the same dollar, and the record prices it on relationships rather than results: the stock’s biggest one-day gain, 32.52%, came after Jensen Huang’s remarks about the data-centre buildout, and in August it expanded its Google chip deal with a warrant of up to $12.2B attached. Custom accelerators are the buyer’s other route around Nvidia’s share of the split — the same instinct as designing memory out, aimed one layer lower. AWS’s plan to add 2M Nvidia GPUs across 2027–28 says the route is a supplement, not yet a replacement.

Micron is the layer currently winning, and the memory wall is where this site keeps that argument in full — including what would break it, which is memory’s own thirty-year history of mean reversion. Read the two pages together: that one asks whether the shortage holds, this one asks who pays for it while it does.

The Lenses

Hamilton Helmer supplies the sentence this whole page is a test of. Invention, he writes in 7 Powers, drives a favourable change in system economics — you get more for less — and the resulting gain will be split somehow between your company and other segments of the value chain. Power is what determines your share of it. The AI buildout is the largest such gain in decades, and the record above is the split being negotiated in public, by press release and price letter. What Helmer’s framework predicts, and the numbers confirm, is that the share does not go to the best technology or the deepest moat. It goes to whoever is scarce at the moment the money is spent. ASML has the better barrier; the memory makers had the better quarter.

Robert Shiller’s narrative economics explains why this keeps surprising people. At any time the market carries one story about what constrains AI — it was chips, then power, then memory — and that story is where attention, and multiple, concentrate. Narratives spread by contagion, not by accounting, which is why the constraint everyone is naming tends to lag the constraint that is actually binding. The rotation is the thing to track; the current name for it is the thing most likely to be already priced. The useful question is never “what is the bottleneck” but “what did the last one cost, and who paid”.

What Moved

Four price increases in four months, aimed up the chain in sequence, is not four companies each having a good year. It is one dollar being re-divided while the customer at the end of it has not yet said no. What to watch is the first link that fails to make its increase stick — TSMC’s resistance to ASML is the live test — because the layer that cannot pass its costs on is the layer that discovers what its power was actually worth.

Sources

Growth figures compare each company’s most recently reported quarter against the same quarter a year earlier, as filed or reported: Samsung and SK Hynix on operating profit, Micron and ASML on revenue, Nvidia and TSMC on both. They are not adjusted for accounting differences between reporting regimes, and a memory maker’s operating profit is a more volatile series than an equipment maker’s revenue by construction — the ladder is a ranking of this cycle’s increment, not a valuation. Article IDs resolve at texxr.com/<id>.